Datasets:

Languages:
English
ArXiv:
License:
FormalArXiv / README.md
Theohhh's picture
Note August 2026 paper source
962166f verified
|
Raw History Blame Contribute Delete
2.94 kB
metadata
pretty_name: FormalArXiv
license: other
language:
  - en
task_categories:
  - text-generation
tags:
  - lean4
  - mathlib
  - theorem-proving
  - formal-mathematics
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl

FormalArXiv

FormalArXiv formalizes in Lean 4 89 research-level statements from papers posted to arXiv in August 2026, sourced from LemmaBench [1].

Authors: Theodore Meek, Michael Zeng, Simon Kurgan, and Pei Li
Institution: University of Washington Math AI Lab

Each JSONL row contains the source-paper metadata, LaTeX statement and assumptions, complete Lean 4 code, main theorem name, and a SHA-256 hash. The main theorem's sorry is the proof-generation target.

from datasets import load_dataset

dataset = load_dataset("uw-math-ai/FormalArXiv", split="test")
print(dataset[0]["lean4_code"])

Benchmark protocol

  • Give the system the natural-language statement, assumptions, and complete Lean code.
  • Replace the single sorry with a proof; do not change the theorem statement, assumptions, definitions, or imports.
  • Allow at most two hours per problem, using one agent and no subagents.
  • Disable internet access. Local Mathlib/LSP search is allowed.
  • Allow unlimited Lean executions within the two-hour limit.
  • Count a problem as solved only when Lean accepts the file with no sorry, admit, or added axioms.
  • Record the system and model version, Lean/Mathlib version, result, and elapsed time for every problem.
  • Report success rate as the number solved out of 89 under the same time and tool budget.
  • Disable provider training or model-improvement use of benchmark data during evaluation.
System Solved Success rate
Aristotle 41 / 89 46.07%
Astra Extra-High 65 / 89 73.03%

Citation

@misc{meek2026formalarxiv,
  author       = {Meek, Theodore and Zeng, Michael and Kurgan, Simon and Li, Pei},
  title        = {{FormalArXiv}: A Formalized Research-level Mathematics Benchmark},
  year         = {2026},
  howpublished = {University of Washington Math AI Lab}
}

See LICENSE for licensing details and source-material attribution.

References

Parts of our formalization methodology build on Meta's Automatic Textbook Formalization [2] and Formalizing Numerical Analysis [3].

  1. Antoine Peyronnet, Fabian Gloeckle, and Amaury Hayat. LemmaBench: A Live, Research-Level Benchmark to Evaluate LLM Capabilities in Mathematics, 2026.
  2. Fabian Gloeckle, Ahmad Rammal, Charles Arnal, Remi Munos, Vivien Cabannes, Gabriel Synnaeve, and Amaury Hayat. Automatic Textbook Formalization, 2026.
  3. Theodore Meek, Siyuan Ge, Di Qiu Xiang, Simon Chess, and Vasily Ilin. Formalizing Numerical Analysis: An Agent Pipeline and Quality Audit Beyond Kernel Acceptance, 2026.